As usual methods extract few features in low contrast image and obtain wrong matching result easily, an improved NCC registration based on self-similarity feature was proposed. The method detects feature point by calculating self-similarity value, then constructs the adaptive window according to the scale information of feature point for matching, and improves matching results by threshold filtering and RANSAC algorithm. The experimental results show that the proposed method possesses higher efficiency than usual algorithm in low contrast images, and it has strong robustness to image scale and rotation transformation.